A Mixture-of-Gradient-Experts Framework for Accelerating AC Optimal Power Flow

Abstract: This paper introduces a novel data-driven constraint screening approach aimed at accelerating the solution of convexified AC optimal power flow (C-OPF) by eliminating non-binding constraints. Our constraint screening process leverages a novel mixture-of-experts architecture called MoGE (Mixture of Gradient Experts), which is trained to predict optimal dual variables based on problem parameters. The results demonstrate that subject to certain mild conditions on the C-OPF model, our proposed method guarantees an identical solution to the full problem but significantly reduces computational time. The solution's accuracy is shown through C-OPF properties. Even with incorrectly screened constraints, a recovery process is possible. This results in a problem with fewer constraints than the original. Extensive simulations conducted on the quadratic-convex model show that our method outperforms other constraint screening techniques.